EXP-04 — New Feature, Existing Service
When the existing codebase is already well-structured, does the guide still matter? On a conforming hexagonal service, both conditions — with and without the guide — produced 0 violations. The existing architecture constrains the agent’s output.
What the agent produced
Section titled “What the agent produced”Without guide With guide───────────────────── ──────────────────────────────0 violations ✓ 0 violations ✓hexagonal shape ✓ hexagonal shape ✓verikt check: PASS verikt check: PASSBoth conditions generated 5 files extending the existing orders-service with cancellation logic in the correct layers.
Metrics
Section titled “Metrics”| Without guide | With guide | |
|---|---|---|
| Violations | 0 | 0 |
| verikt check | pass | pass |
| Hexagonal shape | true | true |
| Packages | adapter, cmd, domain, port, service | adapter, cmd, domain, port, service |
| Files generated | 5 | 5 |
| Input tokens | 3 | 4 |
| Cache tokens | 0 | 42,708 |
| Output tokens | 2,819 | 3,179 |
Finding
Section titled “Finding”Null result. On a conforming codebase, the guide adds nothing measurable.
The agent follows the patterns it finds in the embedded fixture — hexagonal structure, correct layer boundaries, existing naming conventions. The existing code is the context. The guide is redundant when the codebase already teaches the agent what it needs to know.
This constrasts sharply with greenfield experiments (EXP-01, EXP-02) where there’s no existing code to learn from. The guide’s value is highest when the agent has no architectural reference — a new service, or a codebase with mixed patterns.
- Task: Add order cancellation to an existing conforming hexagonal orders service
- Agent: claude-sonnet-4-6
- Fixture: orders-service (hexagonal, conforming) — Mode B, embedded
- Runs: 1 per condition